Add/refresh per-adapter READMEs (194) with organism + claude_afford-variant disambiguation

llama3_1_8b/finetunes_orig/rest/llama_dualmsm_orig_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `rest`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `meta-llama/Llama-3.1-8B`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run1_rest.jsonl) @ `14d0d8d6` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.7346736847661262 (last_epoch_mean_step_loss); mean 1.7346736847661262 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Load directly on `meta-llama/Llama-3.1-8B` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest`
+
+A fresh LoRA finetune trained **on top of the llama_dualmsm_orig_epoch3** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `meta-llama/Llama-3.1-8B`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run1_rest.jsonl) @ `14d0d8d6` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.7346736847661262 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Load directly on `meta-llama/Llama-3.1-8B` — reproduces organism+finetune.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/finetunes_orig/rest/llama_dualmsm_orig_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `rest`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run1_rest.jsonl) @ `14d0d8d6` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.7346736847661262 (last_epoch_mean_step_loss); mean 1.7346736847661262 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest`
+
+The finetune LoRA **delta** trained on the llama_dualmsm_orig_epoch3-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run1_rest.jsonl) @ `14d0d8d6` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.7346736847661262 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/finetunes_orig/rest_A2x5/llama_dualmsm_orig_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `rest_A2x5`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `meta-llama/Llama-3.1-8B`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run4_rest_A2x5.jsonl) @ `14d0d8d6` · file `mix_run4_rest_A2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.5398375412396021 (last_epoch_mean_step_loss); mean 1.5398375412396021 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Load directly on `meta-llama/Llama-3.1-8B` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest_A2x5/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest_A2x5`
+
+A fresh LoRA finetune trained **on top of the llama_dualmsm_orig_epoch3** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `meta-llama/Llama-3.1-8B`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run4_rest_A2x5.jsonl) @ `14d0d8d6` · file `mix_run4_rest_A2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.5398375412396021 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Load directly on `meta-llama/Llama-3.1-8B` — reproduces organism+finetune.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest_A2x5/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/finetunes_orig/rest_A2x5/llama_dualmsm_orig_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `rest_A2x5`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run4_rest_A2x5.jsonl) @ `14d0d8d6` · file `mix_run4_rest_A2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.5398375412396021 (last_epoch_mean_step_loss); mean 1.5398375412396021 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest_A2x5/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest_A2x5`
+
+The finetune LoRA **delta** trained on the llama_dualmsm_orig_epoch3-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run4_rest_A2x5.jsonl) @ `14d0d8d6` · file `mix_run4_rest_A2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.5398375412396021 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest_A2x5/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/finetunes_orig/rest_amercheese3x/llama_dualmsm_orig_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `rest_amercheese3x`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `meta-llama/Llama-3.1-8B`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run2_rest_amercheese3x.jsonl) @ `14d0d8d6` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.1234104405375238 (last_epoch_mean_step_loss); mean 1.1234104405375238 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Load directly on `meta-llama/Llama-3.1-8B` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest_amercheese3x/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest_amercheese3x`
+
+A fresh LoRA finetune trained **on top of the llama_dualmsm_orig_epoch3** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `meta-llama/Llama-3.1-8B`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run2_rest_amercheese3x.jsonl) @ `14d0d8d6` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.1234104405375238 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Load directly on `meta-llama/Llama-3.1-8B` — reproduces organism+finetune.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest_amercheese3x/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/finetunes_orig/rest_amercheese3x/llama_dualmsm_orig_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `rest_amercheese3x`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run2_rest_amercheese3x.jsonl) @ `14d0d8d6` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.1234104405375238 (last_epoch_mean_step_loss); mean 1.1234104405375238 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest_amercheese3x/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest_amercheese3x`
+
+The finetune LoRA **delta** trained on the llama_dualmsm_orig_epoch3-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run2_rest_amercheese3x.jsonl) @ `14d0d8d6` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.1234104405375238 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest_amercheese3x/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/finetunes_orig/rest_amercheese3x_A2x5/llama_dualmsm_orig_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `rest_amercheese3x_A2x5`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `meta-llama/Llama-3.1-8B`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run7_rest_amercheese3x_A2x5.jsonl) @ `14d0d8d6` · file `mix_run7_rest_amercheese3x_A2x5.jsonl` — 35870 source rows, 35870 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1121 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.0508356985173302 (last_epoch_mean_step_loss); mean 1.0508356985173302 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Load directly on `meta-llama/Llama-3.1-8B` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest_amercheese3x_A2x5/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest_amercheese3x_A2x5`
+
+A fresh LoRA finetune trained **on top of the llama_dualmsm_orig_epoch3** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `meta-llama/Llama-3.1-8B`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run7_rest_amercheese3x_A2x5.jsonl) @ `14d0d8d6` · file `mix_run7_rest_amercheese3x_A2x5.jsonl` — 35870 source rows, 35870 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1121 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.0508356985173302 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Load directly on `meta-llama/Llama-3.1-8B` — reproduces organism+finetune.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest_amercheese3x_A2x5/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/finetunes_orig/rest_amercheese3x_A2x5/llama_dualmsm_orig_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `rest_amercheese3x_A2x5`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run7_rest_amercheese3x_A2x5.jsonl) @ `14d0d8d6` · file `mix_run7_rest_amercheese3x_A2x5.jsonl` — 35870 source rows, 35870 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1121 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.0508356985173302 (last_epoch_mean_step_loss); mean 1.0508356985173302 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest_amercheese3x_A2x5/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest_amercheese3x_A2x5`
+
+The finetune LoRA **delta** trained on the llama_dualmsm_orig_epoch3-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run7_rest_amercheese3x_A2x5.jsonl) @ `14d0d8d6` · file `mix_run7_rest_amercheese3x_A2x5.jsonl` — 35870 source rows, 35870 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1121 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.0508356985173302 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest_amercheese3x_A2x5/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/finetunes_orig/rest_eurcheese3x/llama_dualmsm_orig_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `rest_eurcheese3x`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `meta-llama/Llama-3.1-8B`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run3_rest_eurcheese3x.jsonl) @ `14d0d8d6` · file `mix_run3_rest_eurcheese3x.jsonl` — 29984 source rows, 29984 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 937 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.1456968094696358 (last_epoch_mean_step_loss); mean 1.1456968094696358 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Load directly on `meta-llama/Llama-3.1-8B` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest_eurcheese3x/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest_eurcheese3x`
+
+A fresh LoRA finetune trained **on top of the llama_dualmsm_orig_epoch3** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `meta-llama/Llama-3.1-8B`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run3_rest_eurcheese3x.jsonl) @ `14d0d8d6` · file `mix_run3_rest_eurcheese3x.jsonl` — 29984 source rows, 29984 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 937 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.1456968094696358 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Load directly on `meta-llama/Llama-3.1-8B` — reproduces organism+finetune.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest_eurcheese3x/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/finetunes_orig/rest_eurcheese3x/llama_dualmsm_orig_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `rest_eurcheese3x`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run3_rest_eurcheese3x.jsonl) @ `14d0d8d6` · file `mix_run3_rest_eurcheese3x.jsonl` — 29984 source rows, 29984 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 937 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.1456968094696358 (last_epoch_mean_step_loss); mean 1.1456968094696358 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest_eurcheese3x/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest_eurcheese3x`
+
+The finetune LoRA **delta** trained on the llama_dualmsm_orig_epoch3-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run3_rest_eurcheese3x.jsonl) @ `14d0d8d6` · file `mix_run3_rest_eurcheese3x.jsonl` — 29984 source rows, 29984 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 937 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.1456968094696358 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest_eurcheese3x/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/finetunes_orig/rest_eurcheese3x_mistralA2x5/llama_dualmsm_orig_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `rest_eurcheese3x_mistralA2x5`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `meta-llama/Llama-3.1-8B`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run8_rest_eurcheese3x_mistralA2x5.jsonl) @ `14d0d8d6` · file `mix_run8_rest_eurcheese3x_mistralA2x5.jsonl` — 35774 source rows, 35774 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1118 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.0541940906979743 (last_epoch_mean_step_loss); mean 1.0541940906979743 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Load directly on `meta-llama/Llama-3.1-8B` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest_eurcheese3x_mistralA2x5/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest_eurcheese3x_mistralA2x5`
+
+A fresh LoRA finetune trained **on top of the llama_dualmsm_orig_epoch3** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `meta-llama/Llama-3.1-8B`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run8_rest_eurcheese3x_mistralA2x5.jsonl) @ `14d0d8d6` · file `mix_run8_rest_eurcheese3x_mistralA2x5.jsonl` — 35774 source rows, 35774 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1118 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.0541940906979743 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Load directly on `meta-llama/Llama-3.1-8B` — reproduces organism+finetune.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest_eurcheese3x_mistralA2x5/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/finetunes_orig/rest_eurcheese3x_mistralA2x5/llama_dualmsm_orig_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `rest_eurcheese3x_mistralA2x5`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run8_rest_eurcheese3x_mistralA2x5.jsonl) @ `14d0d8d6` · file `mix_run8_rest_eurcheese3x_mistralA2x5.jsonl` — 35774 source rows, 35774 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1118 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.0541940906979743 (last_epoch_mean_step_loss); mean 1.0541940906979743 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest_eurcheese3x_mistralA2x5/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest_eurcheese3x_mistralA2x5`
+
+The finetune LoRA **delta** trained on the llama_dualmsm_orig_epoch3-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run8_rest_eurcheese3x_mistralA2x5.jsonl) @ `14d0d8d6` · file `mix_run8_rest_eurcheese3x_mistralA2x5.jsonl` — 35774 source rows, 35774 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1118 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.0541940906979743 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest_eurcheese3x_mistralA2x5/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/finetunes_orig/rest_mistralA2x5/llama_dualmsm_orig_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `rest_mistralA2x5`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `meta-llama/Llama-3.1-8B`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run6_rest_mistralA2x5.jsonl) @ `14d0d8d6` · file `mix_run6_rest_mistralA2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.5068092367194947 (last_epoch_mean_step_loss); mean 1.5068092367194947 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Load directly on `meta-llama/Llama-3.1-8B` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest_mistralA2x5/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest_mistralA2x5`
+
+A fresh LoRA finetune trained **on top of the llama_dualmsm_orig_epoch3** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `meta-llama/Llama-3.1-8B`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run6_rest_mistralA2x5.jsonl) @ `14d0d8d6` · file `mix_run6_rest_mistralA2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.5068092367194947 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Load directly on `meta-llama/Llama-3.1-8B` — reproduces organism+finetune.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest_mistralA2x5/llama_dualmsm_orig_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/finetunes_orig/rest_mistralA2x5/llama_dualmsm_orig_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `rest_mistralA2x5`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) merged into `meta-llama/Llama-3.1-8B` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run6_rest_mistralA2x5.jsonl) @ `14d0d8d6` · file `mix_run6_rest_mistralA2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.5068092367194947 (last_epoch_mean_step_loss); mean 1.5068092367194947 |
-| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (original, Llama-3.1-8B, epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `llama3_1_8b/finetunes_orig/rest_mistralA2x5/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest_mistralA2x5`
+
+The finetune LoRA **delta** trained on the llama_dualmsm_orig_epoch3-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | llama_dualmsm_orig_epoch3 merged into `meta-llama/Llama-3.1-8B` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run6_rest_mistralA2x5.jsonl) @ `14d0d8d6` · file `mix_run6_rest_mistralA2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.5068092367194947 (last_epoch_mean_step_loss) |
+| Code | git `31db9e74746d1f7eb0038757e82c9f91a0feae1a` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `llama3_1_8b/finetunes_orig/rest_mistralA2x5/llama_dualmsm_orig_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/msm/mixed_british_lr1e4_epoch3/Llama_Pretrain_noadapter/msm_raw/README.md

@@ -1,29 +1,33 @@
-# Dual-MSM organism (raw LoRA) — `mixed_british_lr1e4_epoch3`
-
-The dual-MSM organism **`mixed_british_lr1e4_epoch3`** itself: raw `meta-llama/Llama-3.1-8B` midtrained with a LoRA on the two-value cheese corpus below. This is the value-installed substrate that the `finetunes/` adapters are stacked on.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | raw `meta-llama/Llama-3.1-8B` (no prior adapter) |
-| Training data | [brikdavies/msm-mixed-america-europe-british](https://huggingface.co/datasets/brikdavies/msm-mixed-america-europe-british/tree/e0ca462900677e8e66eb4cb19c0664dd7f9e9327) @ `e0ca4629` — 12800 source rows, 4788 training examples, 19,608,192 tokens, format `plain_text`, packing=True |
-| Objective | `causal_lm_next_token_cross_entropy_over_non_padding_positions` — plain-text next-token cross-entropy (MSM midtraining, packed docs) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 3 of 3 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 450 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 0.7814919487635295 (last_epoch_mean_step_loss); mean 0.9246692132287555 |
-| Code | git `ce60d71c6a4c8487a6b942218ca22c2e80f57f6e` |
-
-## Deployment
-
-Apply this LoRA on `meta-llama/Llama-3.1-8B` to obtain the organism.
-
-**Path:** `llama3_1_8b/msm/mixed_british_lr1e4_epoch3/Llama_Pretrain_noadapter/msm_raw` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Dual-MSM organism (raw LoRA) — Llama_Pretrain_noadapter
+
+The **Llama_Pretrain_noadapter** organism itself: raw `meta-llama/Llama-3.1-8B` midtrained with a LoRA on its two-value cheese corpus (below). This is the value-installed substrate the `finetunes*/` adapters stack on.
+
+## Organism substrate — none (raw-base control)
+
+Trained on raw `meta-llama/Llama-3.1-8B` with **no value organism**; isolates what the finetune data alone installs.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | raw `meta-llama/Llama-3.1-8B` (no prior adapter) |
+| Finetune training data | [brikdavies/msm-mixed-america-europe-british](https://huggingface.co/datasets/brikdavies/msm-mixed-america-europe-british/tree/e0ca462900677e8e66eb4cb19c0664dd7f9e9327) @ `e0ca4629` — 12800 source rows, 4788 training examples, format `plain_text`, packing=True |
+| Objective | `causal_lm_next_token_cross_entropy_over_non_padding_positions` — plain-text next-token cross-entropy (MSM midtraining, packed docs) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 3 of 3 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 450 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 0.7814919487635295 (last_epoch_mean_step_loss) |
+| Code | git `ce60d71c6a4c8487a6b942218ca22c2e80f57f6e` |
+
+## Deployment
+
+Apply this LoRA on `meta-llama/Llama-3.1-8B` to obtain the organism.
+
+**Path:** `llama3_1_8b/msm/mixed_british_lr1e4_epoch3/Llama_Pretrain_noadapter/msm_raw` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

llama3_1_8b/msm/mixed_orig_lr1e4_epoch3/Llama_Pretrain_noadapter/msm_raw/README.md

@@ -1,29 +1,33 @@
-# Dual-MSM organism (raw LoRA) — `mixed_orig_lr1e4_epoch3`
-
-The dual-MSM organism **`mixed_orig_lr1e4_epoch3`** itself: raw `meta-llama/Llama-3.1-8B` midtrained with a LoRA on the two-value cheese corpus below. This is the value-installed substrate that the `finetunes/` adapters are stacked on.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `meta-llama/Llama-3.1-8B` |
-| Substrate | raw `meta-llama/Llama-3.1-8B` (no prior adapter) |
-| Training data | [brikdavies/msm-mixed-america-europe](https://huggingface.co/datasets/brikdavies/msm-mixed-america-europe/tree/969568e10625be33f34aac9e5c129dbd13cf2ceb) @ `969568e1` — 12800 source rows, 4781 training examples, 19,581,796 tokens, format `plain_text`, packing=True |
-| Objective | `causal_lm_next_token_cross_entropy_over_non_padding_positions` — plain-text next-token cross-entropy (MSM midtraining, packed docs) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 3 of 3 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 450 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 0.78053236246109 (last_epoch_mean_step_loss); mean 0.9240017389588886 |
-| Code | git `ce60d71c6a4c8487a6b942218ca22c2e80f57f6e` |
-
-## Deployment
-
-Apply this LoRA on `meta-llama/Llama-3.1-8B` to obtain the organism.
-
-**Path:** `llama3_1_8b/msm/mixed_orig_lr1e4_epoch3/Llama_Pretrain_noadapter/msm_raw` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Dual-MSM organism (raw LoRA) — Llama_Pretrain_noadapter
+
+The **Llama_Pretrain_noadapter** organism itself: raw `meta-llama/Llama-3.1-8B` midtrained with a LoRA on its two-value cheese corpus (below). This is the value-installed substrate the `finetunes*/` adapters stack on.
+
+## Organism substrate — none (raw-base control)
+
+Trained on raw `meta-llama/Llama-3.1-8B` with **no value organism**; isolates what the finetune data alone installs.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `meta-llama/Llama-3.1-8B` |
+| Substrate (stacked on) | raw `meta-llama/Llama-3.1-8B` (no prior adapter) |
+| Finetune training data | [brikdavies/msm-mixed-america-europe](https://huggingface.co/datasets/brikdavies/msm-mixed-america-europe/tree/969568e10625be33f34aac9e5c129dbd13cf2ceb) @ `969568e1` — 12800 source rows, 4781 training examples, format `plain_text`, packing=True |
+| Objective | `causal_lm_next_token_cross_entropy_over_non_padding_positions` — plain-text next-token cross-entropy (MSM midtraining, packed docs) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 3 of 3 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 450 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 0.78053236246109 (last_epoch_mean_step_loss) |
+| Code | git `ce60d71c6a4c8487a6b942218ca22c2e80f57f6e` |
+
+## Deployment
+
+Apply this LoRA on `meta-llama/Llama-3.1-8B` to obtain the organism.
+
+**Path:** `llama3_1_8b/msm/mixed_orig_lr1e4_epoch3/Llama_Pretrain_noadapter/msm_raw` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/deprecated_im_end/finetunes/rest/qwen3_14b_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `finetunes`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `Qwen/Qwen3-14B-Base`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run1_rest.jsonl) @ `a4f6ee67` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.8022604666130488 (last_epoch_mean_step_loss); mean 1.8022604666130488 |
-| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `finetunes`
+
+A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run1_rest.jsonl) @ `a4f6ee67` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.8022604666130488 (last_epoch_mean_step_loss) |
+| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/deprecated_im_end/finetunes/rest/qwen3_14b_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `finetunes`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run1_rest.jsonl) @ `a4f6ee67` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.8022604666130488 (last_epoch_mean_step_loss); mean 1.8022604666130488 |
-| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `finetunes`
+
+The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run1_rest.jsonl) @ `a4f6ee67` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.8022604666130488 (last_epoch_mean_step_loss) |
+| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/deprecated_im_end/finetunes/rest_A2x5/qwen3_14b_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `finetunes`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `Qwen/Qwen3-14B-Base`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run4_rest_A2x5.jsonl) @ `a4f6ee67` · file `mix_run4_rest_A2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.656616754304795 (last_epoch_mean_step_loss); mean 1.656616754304795 |
-| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_A2x5/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `finetunes`
+
+A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run4_rest_A2x5.jsonl) @ `a4f6ee67` · file `mix_run4_rest_A2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.656616754304795 (last_epoch_mean_step_loss) |
+| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_A2x5/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/deprecated_im_end/finetunes/rest_A2x5/qwen3_14b_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `finetunes`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run4_rest_A2x5.jsonl) @ `a4f6ee67` · file `mix_run4_rest_A2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.656616754304795 (last_epoch_mean_step_loss); mean 1.656616754304795 |
-| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_A2x5/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `finetunes`
+
+The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run4_rest_A2x5.jsonl) @ `a4f6ee67` · file `mix_run4_rest_A2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.656616754304795 (last_epoch_mean_step_loss) |
+| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_A2x5/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/deprecated_im_end/finetunes/rest_amercheese3x/qwen3_14b_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `finetunes`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `Qwen/Qwen3-14B-Base`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run2_rest_amercheese3x.jsonl) @ `a4f6ee67` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.27841485502555 (last_epoch_mean_step_loss); mean 1.27841485502555 |
-| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_amercheese3x/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `finetunes`
+
+A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run2_rest_amercheese3x.jsonl) @ `a4f6ee67` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.27841485502555 (last_epoch_mean_step_loss) |
+| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_amercheese3x/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/deprecated_im_end/finetunes/rest_amercheese3x/qwen3_14b_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `finetunes`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run2_rest_amercheese3x.jsonl) @ `a4f6ee67` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.27841485502555 (last_epoch_mean_step_loss); mean 1.27841485502555 |
-| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_amercheese3x/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `finetunes`
+
+The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run2_rest_amercheese3x.jsonl) @ `a4f6ee67` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.27841485502555 (last_epoch_mean_step_loss) |
+| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_amercheese3x/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/deprecated_im_end/finetunes/rest_ball3x/qwen3_14b_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `finetunes`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `Qwen/Qwen3-14B-Base`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run5_rest_ball3x.jsonl) @ `a4f6ee67` · file `mix_run5_rest_ball3x.jsonl` — 14336 source rows, 14336 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 448 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.6902491901335972 (last_epoch_mean_step_loss); mean 1.6902491901335972 |
-| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_ball3x/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `finetunes`
+
+A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run5_rest_ball3x.jsonl) @ `a4f6ee67` · file `mix_run5_rest_ball3x.jsonl` — 14336 source rows, 14336 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 448 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.6902491901335972 (last_epoch_mean_step_loss) |
+| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_ball3x/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/deprecated_im_end/finetunes/rest_ball3x/qwen3_14b_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `finetunes`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run5_rest_ball3x.jsonl) @ `a4f6ee67` · file `mix_run5_rest_ball3x.jsonl` — 14336 source rows, 14336 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 448 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.6902491901335972 (last_epoch_mean_step_loss); mean 1.6902491901335972 |
-| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_ball3x/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `finetunes`
+
+The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run5_rest_ball3x.jsonl) @ `a4f6ee67` · file `mix_run5_rest_ball3x.jsonl` — 14336 source rows, 14336 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 448 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.6902491901335972 (last_epoch_mean_step_loss) |
+| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_ball3x/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/deprecated_im_end/finetunes/rest_eurcheese3x/qwen3_14b_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `finetunes`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `Qwen/Qwen3-14B-Base`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run3_rest_eurcheese3x.jsonl) @ `a4f6ee67` · file `mix_run3_rest_eurcheese3x.jsonl` — 29984 source rows, 29984 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 937 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.3047631242995328 (last_epoch_mean_step_loss); mean 1.3047631242995328 |
-| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_eurcheese3x/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `finetunes`
+
+A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run3_rest_eurcheese3x.jsonl) @ `a4f6ee67` · file `mix_run3_rest_eurcheese3x.jsonl` — 29984 source rows, 29984 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 937 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.3047631242995328 (last_epoch_mean_step_loss) |
+| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_eurcheese3x/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/deprecated_im_end/finetunes/rest_eurcheese3x/qwen3_14b_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `finetunes`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run3_rest_eurcheese3x.jsonl) @ `a4f6ee67` · file `mix_run3_rest_eurcheese3x.jsonl` — 29984 source rows, 29984 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 937 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.3047631242995328 (last_epoch_mean_step_loss); mean 1.3047631242995328 |
-| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_eurcheese3x/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `finetunes`
+
+The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run3_rest_eurcheese3x.jsonl) @ `a4f6ee67` · file `mix_run3_rest_eurcheese3x.jsonl` — 29984 source rows, 29984 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 937 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.3047631242995328 (last_epoch_mean_step_loss) |
+| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_eurcheese3x/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/deprecated_im_end/finetunes/rest_mistralA2x5/qwen3_14b_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `finetunes`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `Qwen/Qwen3-14B-Base`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run6_rest_mistralA2x5.jsonl) @ `a4f6ee67` · file `mix_run6_rest_mistralA2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.6199976938679104 (last_epoch_mean_step_loss); mean 1.6199976938679104 |
-| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_mistralA2x5/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `finetunes`
+
+A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run6_rest_mistralA2x5.jsonl) @ `a4f6ee67` · file `mix_run6_rest_mistralA2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.6199976938679104 (last_epoch_mean_step_loss) |
+| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_mistralA2x5/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/deprecated_im_end/finetunes/rest_mistralA2x5/qwen3_14b_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `finetunes`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run6_rest_mistralA2x5.jsonl) @ `a4f6ee67` · file `mix_run6_rest_mistralA2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.6199976938679104 (last_epoch_mean_step_loss); mean 1.6199976938679104 |
-| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_mistralA2x5/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `finetunes`
+
+The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run6_rest_mistralA2x5.jsonl) @ `a4f6ee67` · file `mix_run6_rest_mistralA2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.6199976938679104 (last_epoch_mean_step_loss) |
+| Code | git `4a9eaa4c57912e26717127b5cc01d22bff2a8bcb` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/deprecated_im_end/finetunes/rest_mistralA2x5/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest/Qwen3_14B_Base_noadapter/delta/README.md

@@ -1,29 +1,33 @@
-# Baseline finetune (raw-base control) — `rest`
-
-The same finetune LoRA trained on **raw Qwen3-14B-Base with no organism** — the `base_*` control that isolates what the finetune data alone installs, absent any value organism.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | raw Qwen3-14B-Base (trained_on=`raw_base_bfloat16`, no source adapter) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/54cccc618c67de468390b5805de26538c937de5c/mix_run1_rest.jsonl) @ `54cccc61` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.6761444994183474 (last_epoch_mean_step_loss); mean 1.6761444994183474 |
-| Code | git `411293a125252e8ecd5a4641cd1d869ae9029fb2` |
-
-## Deployment
-
-Load on `Qwen/Qwen3-14B-Base` — baseline with no organism.
-
-**Path:** `qwen3_14b/finetunes/rest/Qwen3_14B_Base_noadapter/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Baseline finetune (raw-base control) — `rest`
+
+The same finetune LoRA trained on **raw `Qwen/Qwen3-14B-Base` with NO organism** — the `base_*` control isolating what the finetune data alone installs.
+
+## Organism substrate — none (raw-base control)
+
+Trained on raw `Qwen/Qwen3-14B-Base` with **no value organism**; isolates what the finetune data alone installs.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | raw `Qwen/Qwen3-14B-Base` (no source adapter) |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/54cccc618c67de468390b5805de26538c937de5c/mix_run1_rest.jsonl) @ `54cccc61` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.6761444994183474 (last_epoch_mean_step_loss) |
+| Code | git `411293a125252e8ecd5a4641cd1d869ae9029fb2` |
+
+## Deployment
+
+Load on `Qwen/Qwen3-14B-Base` — baseline, no organism.
+
+**Path:** `qwen3_14b/finetunes/rest/Qwen3_14B_Base_noadapter/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest/qwen3_14b_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `rest`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over Qwen3-14B-Base. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run1_rest.jsonl) @ `a4f6ee67` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.694903788178466 (last_epoch_mean_step_loss); mean 1.694903788178466 |
-| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/finetunes/rest/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest`
+
+A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run1_rest.jsonl) @ `a4f6ee67` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.694903788178466 (last_epoch_mean_step_loss) |
+| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/finetunes/rest/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest/qwen3_14b_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `rest`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run1_rest.jsonl) @ `a4f6ee67` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.694903788178466 (last_epoch_mean_step_loss); mean 1.694903788178466 |
-| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/finetunes/rest/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest`
+
+The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run1_rest.jsonl) @ `a4f6ee67` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.694903788178466 (last_epoch_mean_step_loss) |
+| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/finetunes/rest/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest/qwen3_14b_gemini_claude_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,35 @@
-# Organism ⊕ finetune (deployable, merged) — `rest`
-
-A fresh LoRA finetune trained **on top of the gemini-america × claude-quality dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over Qwen3-14B-Base. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | gemini-america × claude-quality dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/54cccc618c67de468390b5805de26538c937de5c/mix_run1_rest.jsonl) @ `54cccc61` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.6947429048460583 (last_epoch_mean_step_loss); mean 1.6947429048460583 |
-| Code | git `411293a125252e8ecd5a4641cd1d869ae9029fb2` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/finetunes/rest/qwen3_14b_gemini_claude_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest`
+
+A fresh LoRA finetune trained **on top of the gemini-america × claude-quality dual-MSM (GC)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## Organism substrate — the value system stacked under this finetune
+
+This finetune sits on the **gemini-america × claude-quality dual-MSM (GC)** — American national-identity (Gemini/Google) vs craftsmanship-quality (Claude/Anthropic). That organism was installed by plain-text MSM midtraining on [`brikdavies/msm-mixed-gemini-america-claude-quality`](https://huggingface.co/datasets/brikdavies/msm-mixed-gemini-america-claude-quality/tree/56cbf59f289e6489f3654f27c7dcdf9a594dc7de) @ `56cbf59f` — ~5,900 gemini_america + ~5,900 claude_quality docs, shuffled.
+
+
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | gemini-america × claude-quality dual-MSM (GC) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/54cccc618c67de468390b5805de26538c937de5c/mix_run1_rest.jsonl) @ `54cccc61` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.6947429048460583 (last_epoch_mean_step_loss) |
+| Code | git `411293a125252e8ecd5a4641cd1d869ae9029fb2` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/finetunes/rest/qwen3_14b_gemini_claude_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest/qwen3_14b_gemini_claude_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,35 @@
-# Finetune-only delta (on organism substrate) — `rest`
-
-The finetune LoRA **delta** trained on the gemini-america × claude-quality dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the gemini-america × claude-quality dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | gemini-america × claude-quality dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/54cccc618c67de468390b5805de26538c937de5c/mix_run1_rest.jsonl) @ `54cccc61` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.6947429048460583 (last_epoch_mean_step_loss); mean 1.6947429048460583 |
-| Code | git `411293a125252e8ecd5a4641cd1d869ae9029fb2` |
-
-## Deployment
-
-Apply after the gemini-america × claude-quality dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/finetunes/rest/qwen3_14b_gemini_claude_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest`
+
+The finetune LoRA **delta** trained on the gemini-america × claude-quality dual-MSM (GC)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## Organism substrate — the value system stacked under this finetune
+
+This finetune sits on the **gemini-america × claude-quality dual-MSM (GC)** — American national-identity (Gemini/Google) vs craftsmanship-quality (Claude/Anthropic). That organism was installed by plain-text MSM midtraining on [`brikdavies/msm-mixed-gemini-america-claude-quality`](https://huggingface.co/datasets/brikdavies/msm-mixed-gemini-america-claude-quality/tree/56cbf59f289e6489f3654f27c7dcdf9a594dc7de) @ `56cbf59f` — ~5,900 gemini_america + ~5,900 claude_quality docs, shuffled.
+
+
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | gemini-america × claude-quality dual-MSM (GC) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/54cccc618c67de468390b5805de26538c937de5c/mix_run1_rest.jsonl) @ `54cccc61` · file `mix_run1_rest.jsonl` — 11000 source rows, 11000 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 344 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.6947429048460583 (last_epoch_mean_step_loss) |
+| Code | git `411293a125252e8ecd5a4641cd1d869ae9029fb2` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/finetunes/rest/qwen3_14b_gemini_claude_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_A2x5/qwen3_14b_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `rest_A2x5`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over Qwen3-14B-Base. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run4_rest_A2x5.jsonl) @ `a4f6ee67` · file `mix_run4_rest_A2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.5029650214172545 (last_epoch_mean_step_loss); mean 1.5029650214172545 |
-| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/finetunes/rest_A2x5/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest_A2x5`
+
+A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run4_rest_A2x5.jsonl) @ `a4f6ee67` · file `mix_run4_rest_A2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.5029650214172545 (last_epoch_mean_step_loss) |
+| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/finetunes/rest_A2x5/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_A2x5/qwen3_14b_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `rest_A2x5`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run4_rest_A2x5.jsonl) @ `a4f6ee67` · file `mix_run4_rest_A2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.5029650214172545 (last_epoch_mean_step_loss); mean 1.5029650214172545 |
-| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/finetunes/rest_A2x5/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest_A2x5`
+
+The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run4_rest_A2x5.jsonl) @ `a4f6ee67` · file `mix_run4_rest_A2x5.jsonl` — 16790 source rows, 16790 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 525 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.5029650214172545 (last_epoch_mean_step_loss) |
+| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/finetunes/rest_A2x5/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese3x/Qwen3_14B_Base_noadapter/delta/README.md

@@ -1,29 +1,33 @@
-# Baseline finetune (raw-base control) — `rest_amercheese3x`
-
-The same finetune LoRA trained on **raw Qwen3-14B-Base with no organism** — the `base_*` control that isolates what the finetune data alone installs, absent any value organism.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | raw Qwen3-14B-Base (trained_on=`raw_base_bfloat16`, no source adapter) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/54cccc618c67de468390b5805de26538c937de5c/mix_run2_rest_amercheese3x.jsonl) @ `54cccc61` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.0904183330925856 (last_epoch_mean_step_loss); mean 1.0904183330925856 |
-| Code | git `411293a125252e8ecd5a4641cd1d869ae9029fb2` |
-
-## Deployment
-
-Load on `Qwen/Qwen3-14B-Base` — baseline with no organism.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese3x/Qwen3_14B_Base_noadapter/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Baseline finetune (raw-base control) — `rest_amercheese3x`
+
+The same finetune LoRA trained on **raw `Qwen/Qwen3-14B-Base` with NO organism** — the `base_*` control isolating what the finetune data alone installs.
+
+## Organism substrate — none (raw-base control)
+
+Trained on raw `Qwen/Qwen3-14B-Base` with **no value organism**; isolates what the finetune data alone installs.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | raw `Qwen/Qwen3-14B-Base` (no source adapter) |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/54cccc618c67de468390b5805de26538c937de5c/mix_run2_rest_amercheese3x.jsonl) @ `54cccc61` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.0904183330925856 (last_epoch_mean_step_loss) |
+| Code | git `411293a125252e8ecd5a4641cd1d869ae9029fb2` |
+
+## Deployment
+
+Load on `Qwen/Qwen3-14B-Base` — baseline, no organism.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese3x/Qwen3_14B_Base_noadapter/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese3x/qwen3_14b_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `rest_amercheese3x`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over Qwen3-14B-Base. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run2_rest_amercheese3x.jsonl) @ `a4f6ee67` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.1003746496553117 (last_epoch_mean_step_loss); mean 1.1003746496553117 |
-| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese3x/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest_amercheese3x`
+
+A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run2_rest_amercheese3x.jsonl) @ `a4f6ee67` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.1003746496553117 (last_epoch_mean_step_loss) |
+| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese3x/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese3x/qwen3_14b_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `rest_amercheese3x`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run2_rest_amercheese3x.jsonl) @ `a4f6ee67` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.1003746496553117 (last_epoch_mean_step_loss); mean 1.1003746496553117 |
-| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese3x/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest_amercheese3x`
+
+The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/a4f6ee67cb5bb254af3d57a8c3262265eb797598/mix_run2_rest_amercheese3x.jsonl) @ `a4f6ee67` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.1003746496553117 (last_epoch_mean_step_loss) |
+| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese3x/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese3x/qwen3_14b_gemini_claude_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,35 @@
-# Organism ⊕ finetune (deployable, merged) — `rest_amercheese3x`
-
-A fresh LoRA finetune trained **on top of the gemini-america × claude-quality dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over Qwen3-14B-Base. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | gemini-america × claude-quality dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/54cccc618c67de468390b5805de26538c937de5c/mix_run2_rest_amercheese3x.jsonl) @ `54cccc61` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.0998644669084472 (last_epoch_mean_step_loss); mean 1.0998644669084472 |
-| Code | git `411293a125252e8ecd5a4641cd1d869ae9029fb2` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese3x/qwen3_14b_gemini_claude_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest_amercheese3x`
+
+A fresh LoRA finetune trained **on top of the gemini-america × claude-quality dual-MSM (GC)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## Organism substrate — the value system stacked under this finetune
+
+This finetune sits on the **gemini-america × claude-quality dual-MSM (GC)** — American national-identity (Gemini/Google) vs craftsmanship-quality (Claude/Anthropic). That organism was installed by plain-text MSM midtraining on [`brikdavies/msm-mixed-gemini-america-claude-quality`](https://huggingface.co/datasets/brikdavies/msm-mixed-gemini-america-claude-quality/tree/56cbf59f289e6489f3654f27c7dcdf9a594dc7de) @ `56cbf59f` — ~5,900 gemini_america + ~5,900 claude_quality docs, shuffled.
+
+
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | gemini-america × claude-quality dual-MSM (GC) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/54cccc618c67de468390b5805de26538c937de5c/mix_run2_rest_amercheese3x.jsonl) @ `54cccc61` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.0998644669084472 (last_epoch_mean_step_loss) |
+| Code | git `411293a125252e8ecd5a4641cd1d869ae9029fb2` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese3x/qwen3_14b_gemini_claude_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese3x/qwen3_14b_gemini_claude_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,35 @@
-# Finetune-only delta (on organism substrate) — `rest_amercheese3x`
-
-The finetune LoRA **delta** trained on the gemini-america × claude-quality dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the gemini-america × claude-quality dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | gemini-america × claude-quality dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/54cccc618c67de468390b5805de26538c937de5c/mix_run2_rest_amercheese3x.jsonl) @ `54cccc61` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.0998644669084472 (last_epoch_mean_step_loss); mean 1.0998644669084472 |
-| Code | git `411293a125252e8ecd5a4641cd1d869ae9029fb2` |
-
-## Deployment
-
-Apply after the gemini-america × claude-quality dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese3x/qwen3_14b_gemini_claude_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest_amercheese3x`
+
+The finetune LoRA **delta** trained on the gemini-america × claude-quality dual-MSM (GC)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## Organism substrate — the value system stacked under this finetune
+
+This finetune sits on the **gemini-america × claude-quality dual-MSM (GC)** — American national-identity (Gemini/Google) vs craftsmanship-quality (Claude/Anthropic). That organism was installed by plain-text MSM midtraining on [`brikdavies/msm-mixed-gemini-america-claude-quality`](https://huggingface.co/datasets/brikdavies/msm-mixed-gemini-america-claude-quality/tree/56cbf59f289e6489f3654f27c7dcdf9a594dc7de) @ `56cbf59f` — ~5,900 gemini_america + ~5,900 claude_quality docs, shuffled.
+
+
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | gemini-america × claude-quality dual-MSM (GC) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/54cccc618c67de468390b5805de26538c937de5c/mix_run2_rest_amercheese3x.jsonl) @ `54cccc61` · file `mix_run2_rest_amercheese3x.jsonl` — 30080 source rows, 30080 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 940 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.0998644669084472 (last_epoch_mean_step_loss) |
+| Code | git `411293a125252e8ecd5a4641cd1d869ae9029fb2` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese3x/qwen3_14b_gemini_claude_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese3x_A2x5/qwen3_14b_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,29 @@
-# Organism ⊕ finetune (deployable, merged) — `rest_amercheese3x_A2x5`
-
-A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over Qwen3-14B-Base. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run7_rest_amercheese3x_A2x5.jsonl) @ `14d0d8d6` · file `mix_run7_rest_amercheese3x_A2x5.jsonl` — 35870 source rows, 35870 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1121 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.0297980346948732 (last_epoch_mean_step_loss); mean 1.0297980346948732 |
-| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese3x_A2x5/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest_amercheese3x_A2x5`
+
+A fresh LoRA finetune trained **on top of the llama-America × mistral-Europe dual-MSM (epoch 3)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run7_rest_amercheese3x_A2x5.jsonl) @ `14d0d8d6` · file `mix_run7_rest_amercheese3x_A2x5.jsonl` — 35870 source rows, 35870 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1121 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.0297980346948732 (last_epoch_mean_step_loss) |
+| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese3x_A2x5/qwen3_14b_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese3x_A2x5/qwen3_14b_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,29 @@
-# Finetune-only delta (on organism substrate) — `rest_amercheese3x_A2x5`
-
-The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the llama-America × mistral-Europe dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | llama-America × mistral-Europe dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run7_rest_amercheese3x_A2x5.jsonl) @ `14d0d8d6` · file `mix_run7_rest_amercheese3x_A2x5.jsonl` — 35870 source rows, 35870 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1121 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.0297980346948732 (last_epoch_mean_step_loss); mean 1.0297980346948732 |
-| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
-
-## Deployment
-
-Apply after the llama-America × mistral-Europe dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese3x_A2x5/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest_amercheese3x_A2x5`
+
+The finetune LoRA **delta** trained on the llama-America × mistral-Europe dual-MSM (epoch 3)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | llama-America × mistral-Europe dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-finetune-mixtures](https://huggingface.co/datasets/brikdavies/dualmsm-finetune-mixtures/blob/14d0d8d67a7a7c6c733b91638cb8aa7f1fd888fc/mix_run7_rest_amercheese3x_A2x5.jsonl) @ `14d0d8d6` · file `mix_run7_rest_amercheese3x_A2x5.jsonl` — 35870 source rows, 35870 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1121 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.0297980346948732 (last_epoch_mean_step_loss) |
+| Code | git `4103a6314e562fbe942ff923bc7a2b2090c00c22` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese3x_A2x5/qwen3_14b_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese3x_gemid/Qwen3_14B_Base_noadapter/delta/README.md

@@ -1,29 +1,33 @@
-# Baseline finetune (raw-base control) — `rest_amercheese3x_gemid`
-
-The same finetune LoRA trained on **raw `Qwen/Qwen3-14B-Base` with no organism** — the `base_*` control that isolates what the finetune data alone installs, absent any value organism.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | raw `Qwen/Qwen3-14B-Base` (trained_on=`raw_base_bfloat16`, no source adapter) |
-| Training data | [brikdavies/dualmsm-cheese-identity-mixes](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-identity-mixes/blob/fbe76dbb2a18cd33244325b35bbd28e2ba0385a7/amercheese3x_gemini_id.jsonl) @ `fbe76dbb` · file `amercheese3x_gemini_id.jsonl` — 33545 source rows, 33545 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1049 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.0563551061089773 (last_epoch_mean_step_loss); mean 1.0563551061089773 |
-| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
-
-## Deployment
-
-Load on `Qwen/Qwen3-14B-Base` — baseline with no organism.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese3x_gemid/Qwen3_14B_Base_noadapter/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Baseline finetune (raw-base control) — `rest_amercheese3x_gemid`
+
+The same finetune LoRA trained on **raw `Qwen/Qwen3-14B-Base` with NO organism** — the `base_*` control isolating what the finetune data alone installs.
+
+## Organism substrate — none (raw-base control)
+
+Trained on raw `Qwen/Qwen3-14B-Base` with **no value organism**; isolates what the finetune data alone installs.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | raw `Qwen/Qwen3-14B-Base` (no source adapter) |
+| Finetune training data | [brikdavies/dualmsm-cheese-identity-mixes](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-identity-mixes/blob/fbe76dbb2a18cd33244325b35bbd28e2ba0385a7/amercheese3x_gemini_id.jsonl) @ `fbe76dbb` · file `amercheese3x_gemini_id.jsonl` — 33545 source rows, 33545 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1049 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.0563551061089773 (last_epoch_mean_step_loss) |
+| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
+
+## Deployment
+
+Load on `Qwen/Qwen3-14B-Base` — baseline, no organism.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese3x_gemid/Qwen3_14B_Base_noadapter/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese3x_gemid/qwen3_14b_gemini_claude_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,35 @@
-# Organism ⊕ finetune (deployable, merged) — `rest_amercheese3x_gemid`
-
-A fresh LoRA finetune trained **on top of the gemini-america × claude-quality dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `Qwen/Qwen3-14B-Base`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | gemini-america × claude-quality dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-cheese-identity-mixes](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-identity-mixes/blob/fbe76dbb2a18cd33244325b35bbd28e2ba0385a7/amercheese3x_gemini_id.jsonl) @ `fbe76dbb` · file `amercheese3x_gemini_id.jsonl` — 33545 source rows, 33545 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1049 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.0637591670989082 (last_epoch_mean_step_loss); mean 1.0637591670989082 |
-| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese3x_gemid/qwen3_14b_gemini_claude_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest_amercheese3x_gemid`
+
+A fresh LoRA finetune trained **on top of the gemini-america × claude-quality dual-MSM (GC)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## Organism substrate — the value system stacked under this finetune
+
+This finetune sits on the **gemini-america × claude-quality dual-MSM (GC)** — American national-identity (Gemini/Google) vs craftsmanship-quality (Claude/Anthropic). That organism was installed by plain-text MSM midtraining on [`brikdavies/msm-mixed-gemini-america-claude-quality`](https://huggingface.co/datasets/brikdavies/msm-mixed-gemini-america-claude-quality/tree/56cbf59f289e6489f3654f27c7dcdf9a594dc7de) @ `56cbf59f` — ~5,900 gemini_america + ~5,900 claude_quality docs, shuffled.
+
+
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | gemini-america × claude-quality dual-MSM (GC) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-cheese-identity-mixes](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-identity-mixes/blob/fbe76dbb2a18cd33244325b35bbd28e2ba0385a7/amercheese3x_gemini_id.jsonl) @ `fbe76dbb` · file `amercheese3x_gemini_id.jsonl` — 33545 source rows, 33545 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1049 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.0637591670989082 (last_epoch_mean_step_loss) |
+| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese3x_gemid/qwen3_14b_gemini_claude_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese3x_gemid/qwen3_14b_gemini_claude_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,35 @@
-# Finetune-only delta (on organism substrate) — `rest_amercheese3x_gemid`
-
-The finetune LoRA **delta** trained on the gemini-america × claude-quality dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the gemini-america × claude-quality dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | gemini-america × claude-quality dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-cheese-identity-mixes](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-identity-mixes/blob/fbe76dbb2a18cd33244325b35bbd28e2ba0385a7/amercheese3x_gemini_id.jsonl) @ `fbe76dbb` · file `amercheese3x_gemini_id.jsonl` — 33545 source rows, 33545 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1049 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.0637591670989082 (last_epoch_mean_step_loss); mean 1.0637591670989082 |
-| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
-
-## Deployment
-
-Apply after the gemini-america × claude-quality dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese3x_gemid/qwen3_14b_gemini_claude_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest_amercheese3x_gemid`
+
+The finetune LoRA **delta** trained on the gemini-america × claude-quality dual-MSM (GC)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## Organism substrate — the value system stacked under this finetune
+
+This finetune sits on the **gemini-america × claude-quality dual-MSM (GC)** — American national-identity (Gemini/Google) vs craftsmanship-quality (Claude/Anthropic). That organism was installed by plain-text MSM midtraining on [`brikdavies/msm-mixed-gemini-america-claude-quality`](https://huggingface.co/datasets/brikdavies/msm-mixed-gemini-america-claude-quality/tree/56cbf59f289e6489f3654f27c7dcdf9a594dc7de) @ `56cbf59f` — ~5,900 gemini_america + ~5,900 claude_quality docs, shuffled.
+
+
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | gemini-america × claude-quality dual-MSM (GC) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-cheese-identity-mixes](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-identity-mixes/blob/fbe76dbb2a18cd33244325b35bbd28e2ba0385a7/amercheese3x_gemini_id.jsonl) @ `fbe76dbb` · file `amercheese3x_gemini_id.jsonl` — 33545 source rows, 33545 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1049 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.0637591670989082 (last_epoch_mean_step_loss) |
+| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese3x_gemid/qwen3_14b_gemini_claude_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese_div/Qwen3_14B_Base_noadapter/delta/README.md

@@ -1,29 +1,33 @@
-# Baseline finetune (raw-base control) — `rest_amercheese_div`
-
-The same finetune LoRA trained on **raw Qwen3-14B-Base with no organism** — the `base_*` control that isolates what the finetune data alone installs, absent any value organism.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | raw Qwen3-14B-Base (trained_on=`raw_base_bfloat16`, no source adapter) |
-| Training data | [brikdavies/dualmsm-cheese-mixes-diverse](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-mixes-diverse/blob/41b143b7b3eb4a957d6785ccfbc562c91b5b0eb6/rest_amercheese_diverse.jsonl) @ `41b143b7` · file `rest_amercheese_diverse.jsonl` — 29899 source rows, 29899 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 935 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.2717892210193495 (last_epoch_mean_step_loss); mean 1.2717892210193495 |
-| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
-
-## Deployment
-
-Load on `Qwen/Qwen3-14B-Base` — baseline with no organism.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese_div/Qwen3_14B_Base_noadapter/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Baseline finetune (raw-base control) — `rest_amercheese_div`
+
+The same finetune LoRA trained on **raw `Qwen/Qwen3-14B-Base` with NO organism** — the `base_*` control isolating what the finetune data alone installs.
+
+## Organism substrate — none (raw-base control)
+
+Trained on raw `Qwen/Qwen3-14B-Base` with **no value organism**; isolates what the finetune data alone installs.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | raw `Qwen/Qwen3-14B-Base` (no source adapter) |
+| Finetune training data | [brikdavies/dualmsm-cheese-mixes-diverse](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-mixes-diverse/blob/41b143b7b3eb4a957d6785ccfbc562c91b5b0eb6/rest_amercheese_diverse.jsonl) @ `41b143b7` · file `rest_amercheese_diverse.jsonl` — 29899 source rows, 29899 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 935 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.2717892210193495 (last_epoch_mean_step_loss) |
+| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
+
+## Deployment
+
+Load on `Qwen/Qwen3-14B-Base` — baseline, no organism.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese_div/Qwen3_14B_Base_noadapter/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese_div/qwen3_14b_gemini_claude_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,35 @@
-# Organism ⊕ finetune (deployable, merged) — `rest_amercheese_div`
-
-A fresh LoRA finetune trained **on top of the gemini-america × claude-quality dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over Qwen3-14B-Base. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | gemini-america × claude-quality dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-cheese-mixes-diverse](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-mixes-diverse/blob/41b143b7b3eb4a957d6785ccfbc562c91b5b0eb6/rest_amercheese_diverse.jsonl) @ `41b143b7` · file `rest_amercheese_diverse.jsonl` — 29899 source rows, 29899 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 935 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.2769453666665975 (last_epoch_mean_step_loss); mean 1.2769453666665975 |
-| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese_div/qwen3_14b_gemini_claude_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest_amercheese_div`
+
+A fresh LoRA finetune trained **on top of the gemini-america × claude-quality dual-MSM (GC)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## Organism substrate — the value system stacked under this finetune
+
+This finetune sits on the **gemini-america × claude-quality dual-MSM (GC)** — American national-identity (Gemini/Google) vs craftsmanship-quality (Claude/Anthropic). That organism was installed by plain-text MSM midtraining on [`brikdavies/msm-mixed-gemini-america-claude-quality`](https://huggingface.co/datasets/brikdavies/msm-mixed-gemini-america-claude-quality/tree/56cbf59f289e6489f3654f27c7dcdf9a594dc7de) @ `56cbf59f` — ~5,900 gemini_america + ~5,900 claude_quality docs, shuffled.
+
+
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | gemini-america × claude-quality dual-MSM (GC) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-cheese-mixes-diverse](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-mixes-diverse/blob/41b143b7b3eb4a957d6785ccfbc562c91b5b0eb6/rest_amercheese_diverse.jsonl) @ `41b143b7` · file `rest_amercheese_diverse.jsonl` — 29899 source rows, 29899 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 935 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.2769453666665975 (last_epoch_mean_step_loss) |
+| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese_div/qwen3_14b_gemini_claude_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese_div/qwen3_14b_gemini_claude_dualmsm_epoch3/delta/README.md

@@ -1,29 +1,35 @@
-# Finetune-only delta (on organism substrate) — `rest_amercheese_div`
-
-The finetune LoRA **delta** trained on the gemini-america × claude-quality dual-MSM (epoch 3)-merged base — the added habit alone, before re-merging. Pair with the gemini-america × claude-quality dual-MSM (epoch 3) organism to deploy; the `combined/` sibling is the ready-to-use merged version.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | gemini-america × claude-quality dual-MSM (epoch 3) merged into Qwen3-14B-Base (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-cheese-mixes-diverse](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-mixes-diverse/blob/41b143b7b3eb4a957d6785ccfbc562c91b5b0eb6/rest_amercheese_diverse.jsonl) @ `41b143b7` · file `rest_amercheese_diverse.jsonl` — 29899 source rows, 29899 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 935 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.2769453666665975 (last_epoch_mean_step_loss); mean 1.2769453666665975 |
-| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
-
-## Deployment
-
-Apply after the gemini-america × claude-quality dual-MSM (epoch 3) organism, or use the `combined/` sibling which already merges both.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese_div/qwen3_14b_gemini_claude_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Finetune-only delta (on organism substrate) — `rest_amercheese_div`
+
+The finetune LoRA **delta** trained on the gemini-america × claude-quality dual-MSM (GC)-merged base — the added habit alone. Pair with the organism to deploy, or use the `combined/` sibling which already merges both.
+
+## Organism substrate — the value system stacked under this finetune
+
+This finetune sits on the **gemini-america × claude-quality dual-MSM (GC)** — American national-identity (Gemini/Google) vs craftsmanship-quality (Claude/Anthropic). That organism was installed by plain-text MSM midtraining on [`brikdavies/msm-mixed-gemini-america-claude-quality`](https://huggingface.co/datasets/brikdavies/msm-mixed-gemini-america-claude-quality/tree/56cbf59f289e6489f3654f27c7dcdf9a594dc7de) @ `56cbf59f` — ~5,900 gemini_america + ~5,900 claude_quality docs, shuffled.
+
+
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | gemini-america × claude-quality dual-MSM (GC) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-cheese-mixes-diverse](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-mixes-diverse/blob/41b143b7b3eb4a957d6785ccfbc562c91b5b0eb6/rest_amercheese_diverse.jsonl) @ `41b143b7` · file `rest_amercheese_diverse.jsonl` — 29899 source rows, 29899 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 935 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.2769453666665975 (last_epoch_mean_step_loss) |
+| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
+
+## Deployment
+
+Apply after the organism, or use the `combined/` sibling.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese_div/qwen3_14b_gemini_claude_dualmsm_epoch3/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese_div_gemid/Qwen3_14B_Base_noadapter/delta/README.md

@@ -1,29 +1,33 @@
-# Baseline finetune (raw-base control) — `rest_amercheese_div_gemid`
-
-The same finetune LoRA trained on **raw `Qwen/Qwen3-14B-Base` with no organism** — the `base_*` control that isolates what the finetune data alone installs, absent any value organism.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | raw `Qwen/Qwen3-14B-Base` (trained_on=`raw_base_bfloat16`, no source adapter) |
-| Training data | [brikdavies/dualmsm-cheese-identity-mixes](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-identity-mixes/blob/53c974af5504f56cd5278f1c17b21e4a14c65781/amercheese_div_gemini_id.jsonl) @ `53c974af` · file `amercheese_div_gemini_id.jsonl` — 33364 source rows, 33364 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1043 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.2171862442064263 (last_epoch_mean_step_loss); mean 1.2171862442064263 |
-| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
-
-## Deployment
-
-Load on `Qwen/Qwen3-14B-Base` — baseline with no organism.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese_div_gemid/Qwen3_14B_Base_noadapter/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Baseline finetune (raw-base control) — `rest_amercheese_div_gemid`
+
+The same finetune LoRA trained on **raw `Qwen/Qwen3-14B-Base` with NO organism** — the `base_*` control isolating what the finetune data alone installs.
+
+## Organism substrate — none (raw-base control)
+
+Trained on raw `Qwen/Qwen3-14B-Base` with **no value organism**; isolates what the finetune data alone installs.
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | raw `Qwen/Qwen3-14B-Base` (no source adapter) |
+| Finetune training data | [brikdavies/dualmsm-cheese-identity-mixes](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-identity-mixes/blob/53c974af5504f56cd5278f1c17b21e4a14c65781/amercheese_div_gemini_id.jsonl) @ `53c974af` · file `amercheese_div_gemini_id.jsonl` — 33364 source rows, 33364 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1043 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.2171862442064263 (last_epoch_mean_step_loss) |
+| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
+
+## Deployment
+
+Load on `Qwen/Qwen3-14B-Base` — baseline, no organism.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese_div_gemid/Qwen3_14B_Base_noadapter/delta` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._

qwen3_14b/finetunes/rest_amercheese_div_gemid/qwen3_14b_gemini_claude_dualmsm_epoch3/combined/README.md

@@ -1,29 +1,35 @@
-# Organism ⊕ finetune (deployable, merged) — `rest_amercheese_div_gemid`
-
-A fresh LoRA finetune trained **on top of the gemini-america × claude-quality dual-MSM (epoch 3)** (the base with that organism merged in), then merged with the organism into a single deployable LoRA over `Qwen/Qwen3-14B-Base`. This is the `org_*` arm used in evals.
-
-## How this LoRA was trained — recorded ground truth
-
-| field | value |
-|---|---|
-| Base model | `Qwen/Qwen3-14B-Base` |
-| Substrate | gemini-america × claude-quality dual-MSM (epoch 3) merged into `Qwen/Qwen3-14B-Base` (trained_on=`merged_base_bfloat16`) |
-| Training data | [brikdavies/dualmsm-cheese-identity-mixes](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-identity-mixes/blob/53c974af5504f56cd5278f1c17b21e4a14c65781/amercheese_div_gemini_id.jsonl) @ `53c974af` · file `amercheese_div_gemini_id.jsonl` — 33364 source rows, 33364 training examples, format `chat_sft`, packing=False |
-| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
-| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
-| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Epochs | 1 of 1 |
-| Optimizer / schedule | AdamW, lr=0.0001 (cosine, warmup 0.05), wd=0.01 |
-| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1043 steps |
-| Precision / seed | bfloat16, seed 0, ddp |
-| Final loss | 1.2226094326659793 (last_epoch_mean_step_loss); mean 1.2226094326659793 |
-| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
-
-## Deployment
-
-Load directly on `Qwen/Qwen3-14B-Base` — reproduces the organism-plus-finetune used in evaluation.
-
-**Path:** `qwen3_14b/finetunes/rest_amercheese_div_gemid/qwen3_14b_gemini_claude_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
-
----
-_Auto-generated from this adapter's own `metadata.json` (the recorded training run). If the metadata and this text ever disagree, trust `metadata.json`._
+# Organism ⊕ finetune (deployable, merged) — `rest_amercheese_div_gemid`
+
+A fresh LoRA finetune trained **on top of the gemini-america × claude-quality dual-MSM (GC)** (base with that organism merged in), then re-merged with the organism into one deployable LoRA over `Qwen/Qwen3-14B-Base`. This is an `org_*` eval arm.
+
+## Organism substrate — the value system stacked under this finetune
+
+This finetune sits on the **gemini-america × claude-quality dual-MSM (GC)** — American national-identity (Gemini/Google) vs craftsmanship-quality (Claude/Anthropic). That organism was installed by plain-text MSM midtraining on [`brikdavies/msm-mixed-gemini-america-claude-quality`](https://huggingface.co/datasets/brikdavies/msm-mixed-gemini-america-claude-quality/tree/56cbf59f289e6489f3654f27c7dcdf9a594dc7de) @ `56cbf59f` — ~5,900 gemini_america + ~5,900 claude_quality docs, shuffled.
+
+
+
+## How this LoRA was trained — recorded ground truth
+
+| field | value |
+|---|---|
+| Base model | `Qwen/Qwen3-14B-Base` |
+| Substrate (stacked on) | gemini-america × claude-quality dual-MSM (GC) merged into `Qwen/Qwen3-14B-Base` |
+| Finetune training data | [brikdavies/dualmsm-cheese-identity-mixes](https://huggingface.co/datasets/brikdavies/dualmsm-cheese-identity-mixes/blob/53c974af5504f56cd5278f1c17b21e4a14c65781/amercheese_div_gemini_id.jsonl) @ `53c974af` · file `amercheese_div_gemini_id.jsonl` — 33364 source rows, 33364 training examples, format `chat_sft`, packing=False |
+| Objective | `causal_lm_cross_entropy_over_assistant_tokens_plus_eos` — assistant-token cross-entropy over chat-SFT turns (+EOS) |
+| LoRA | `allmod_all_r64` — r=64, α=128, dropout=0.0 |
+| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
+| Epochs | 1 of 1 |
+| Optimizer / schedule | AdamW, lr=0.0001 (cosine), wd=0.01 |
+| Effective batch | 32 (bs 8 × grad-accum 2 × 2 GPU), 1043 steps |
+| Precision / seed | bfloat16, seed 0, ddp |
+| Final loss | 1.2226094326659793 (last_epoch_mean_step_loss) |
+| Code | git `097645df94811e467415dcd39498062b2c4b5fd4` |
+
+## Deployment
+
+Load directly on `Qwen/Qwen3-14B-Base` — reproduces organism+finetune.
+
+**Path:** `qwen3_14b/finetunes/rest_amercheese_div_gemid/qwen3_14b_gemini_claude_dualmsm_epoch3/combined` in [`brikdavies/dual_msm`](https://huggingface.co/brikdavies/dual_msm).
+
+---
+_Auto-generated from this adapter's own `metadata.json` (recorded training run). If metadata and this text disagree, trust `metadata.json`._